Kobel compared to every other way of controlling AI access

There is no single market here — there are three, and they are constantly confused with each other. This page maps all of them, says plainly which problem each solves, and tells you when Kobel is the wrong answer.

Last reviewed: 2026-08-19

The landscape at a glance

Every product below is real and worth knowing about. They are grouped by the problem they actually solve, because the biggest source of confusion in this category is people comparing a network proxy to a desktop app.

Three categories that get confused constantly

1. Network-side gateways — governing remote MCP servers

MintMCP, TrueFoundry, Obot, Pomerium, Kong, Cloudflare, IBM ContextForge and Portkey all sit in the network path between an AI client and one or more remote MCP servers. They authenticate the caller against an identity provider, apply role-based rules about which tools that role may invoke, and log everything for compliance. Pomerium's policy language, for example, can match tool names by prefix so that finance may run query tools but not admin tools.

They are designed for the network path, not the endpoint. Several — IBM ContextForge, Obot, agentgateway — are self-hostable and will run perfectly well on a laptop, and TrueFoundry deploys to VPC, on-premise or air-gapped environments. What none of them offer is per-folder permissions over your own filesystem, because that is not the problem they solve. They are bought by IT departments; several publish self-serve tiers, but enterprise pricing is usually on request.

2. Vendor-built controls — governing one assistant

Anthropic, OpenAI and Microsoft all ship permission controls now, and they are getting better fast. This is the category that matters most, because it is free and already installed.

• Claude Desktop and Cowork use attached workspace folders, a manual-versus-automatic approval mode, and always ask before permanently deleting a file. Managed deployments can pin allowed folders to read-only or read-write.
• ChatGPT has the clearest consumer permission tiering in the market — Always ask, Any changes, Important actions, Never ask — but its connectors are cloud-to-cloud OAuth. There is no first-party local filesystem connector; local access goes through user-run MCP servers in developer mode.
• Windows agent workspace gives each agent its own Windows account and asks per agent, with Allow always / Ask every time / Never allow, scoped to six known folders. It is still in preview and Windows-only.

The shared limitation is obvious once you name it: each of these governs only its own assistant. Turning Claude's approvals up does nothing about the second AI tool you installed last month.

3. Local host-side consent — governing your machine

This is where Kobel sits, together with Docker MCP Toolkit, ToolHive Studio, LM Studio, Continue and Cursor. Everything here runs on your computer and decides what an AI may touch before the request reaches the file system. The pattern across the category is striking: almost every one of them does either per-tool gating or per-path scoping, and the ones that do both do it in a YAML file.

Kobel's permission model

Every folder, app and account you add to Kobel gets exactly one of five levels. You set it by dragging a folder in and picking a colour; there is no configuration file.

• GreenRead and write. The AI may open files and save changes in place.
• TealWrite with automatic backup. Every change is preceded by a backup copy, so a bad edit is always reversible.
• OrangeRead the original, write only to copies. The AI can work with your data but cannot alter the source file.
• YellowRead only. Look, don't touch.
• RedBlocked. The resource is invisible to the AI.

Two of these have no equivalent anywhere else in the category. Teal — write, but snapshot first — turns the single most common fear ("what if it overwrites something important?") into an undo button. Orange — read the original, write only copies — is the setting most people actually want for accounting, contracts and client data, and no major AI vendor offers it.

Where Kobel is the strongest option

• You use more than one AI assistant. One ruleset lives on your machine and applies to every connected client — Claude Desktop, ChatGPT Desktop, GitHub Copilot, Gemini CLI, Cursor, Ollama or any other MCP client, auto-detected and connected in one click. Note the trade-off: it is one ruleset. Individual assistants can be switched on or off, not given different rules.
• You want per-folder rules without a config file. Drag, drop, pick a colour.
• You need a gradient, not a switch. Read-only versus read-write is two options. Kobel gives five, including backup-before-write and write-to-copies-only.
• The data must not leave the machine. Kobel evaluates every permission on the device and uploads no file contents.
• You need one log the AI cannot touch. Every file read, change and deletion, every app call and every social post is recorded with timestamp, AI identity and permission level — and the log is not exposed to the AI through the gateway, so a connected assistant can neither read it nor alter it. Exportable as CSV or JSON.
• Files and accounts in one place. Kobel governs local folders, cloud folders synced to disk, in-house software through the App Proxy, and eight social platforms with per-capability limits.

Where Kobel is the wrong tool

Which one should you pick?

Keep reading

Sources